Exploring nexus between particulate pollution and urban land using land use regression (LUR) and machine learning models: a case of study of Delhi, India

Exploring nexus between particulate pollution and urban land using land use regression (LUR) and machine learning models: a case of study of Delhi, India.

Kamna Sachdeva¹ and Divansh Sharma²

¹Professor Department of sustainability sciences, Delhi Skill and Entrepreneurship University (email: kamna.sachdeva@dseu.ac.in)

² Research Fellow, Division of air Quality The energy and Resource Institute (TERI) (email: divyansh.sharma@teri.res.in)

Investigating the environmental repercussions of urban growth dynamics is essential for sustainable urban development. Urbanization affects air pollutants through urban expansion and emission growth, inevitably shifting the health risks associated with air pollution. The interaction between temporal variations of pollutants and spatial heterogeneity further complicates the dynamics of urban air pollution. To cater such heterogeneity regression models are integral they provide detailed insights into the relationships between air pollutants and various influencing factors. These models correlate air pollutants with independent variables, including anthropogenic emissions, meteorological parameters, and the concentrations of other air pollutants. The air quality of Delhi where transboundary emissions, local emissions, land use changes/patters and different seasonal patterns interplays, can only be explained by land use regression models. Land use regression (LUR) modeling, which offers refined insights into the spatial distribution of pollutants by incorporating land use characteristics. The integration of machine learning into land use regression (LUR) modeling further enhance its capability to predict air pollution levels with greater accuracy and spatial resolution. The study was planned to investigate the application of Land Use Regression (LUR) models to explore the relationship between particulate pollution and urban land use in Delhi, incorporating geographic, meteorological, and machine-learning approaches. The study highlights the effectiveness of traditional LUR models, Random Forest (RF), and Deep Neural Networks (DNN) in capturing spatial and temporal variability of PM2.5 and PM10 concentrations. Traditional LUR models were developed for both annual and seasonal predictions, with key variables selected based on their statistical significance and impact direction on pollutant levels. For instance, the annual model for PM2.5 included variables like green cover, building area, and wind speed, while the seasonal models adjusted variables to reflect specific environmental conditions of each period. This methodical selection and modeling process formed the basis for further analysis using advanced techniques. Advanced machine learning models, including RF and DNN, were applied to enhance the traditional LUR models. These models demonstrated improved predictive accuracy and robustness, effectively handling nonlinear interactions and complex data patterns. The study revealed some unexpected trends, particularly in terms of the temporal persistence of pollutants and understanding intensity of pollution hotspots across Delhi.